[WASM SIMD] Replace primitive shuffles with general Shuffle.

src/wasm/function-body-decoder.cc

 // Copyright 2015 the V8 project authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "src/signature.h"
 
 #include "src/base/platform/elapsed-timer.h"
 #include "src/bit-vector.h"
 #include "src/flags.h"
 #include "src/handles.h"
@@ skipping 128 matching lines @@
 
   static Control Try(const byte* pc, size_t stack_depth, SsaEnv* end_env,
                      Zone* zone, SsaEnv* catch_env, int32_t previous_catch) {
     DCHECK_NOT_NULL(catch_env);
     TryInfo* try_info = new (zone) TryInfo(catch_env);
     return {pc,       kControlTry,    stack_depth, end_env,        nullptr,
             try_info, previous_catch, false,       {0, {NO_VALUE}}};
   }
 };
 
+namespace {
+inline unsigned GetShuffleMaskSize(WasmOpcode opcode) {
+  switch (opcode) {
+    case kExprS32x4Shuffle:
+      return 4;
+    case kExprS16x8Shuffle:
+      return 8;
+    case kExprS8x16Shuffle:
+      return 16;
+    default:
+      UNREACHABLE();
+      return 0;
+  }
+}
+}  // namespace
+
 // Macros that build nodes only if there is a graph and the current SSA
 // environment is reachable from start. This avoids problems with malformed
 // TF graphs when decoding inputs that have unreachable code.
 #define BUILD(func, ...) \
   (build() ? CheckForException(builder_->func(__VA_ARGS__)) : nullptr)
 #define BUILD0(func) (build() ? CheckForException(builder_->func()) : nullptr)
 
 // Generic Wasm bytecode decoder with utilities for decoding operands,
 // lengths, etc.
 class WasmDecoder : public Decoder {
@@ skipping 246 matching lines @@
     }
     if (operand.shift < 0 || operand.shift >= max_shift) {
       error(pc_ + 2, "invalid shift amount");
       return false;
     } else {
       return true;
     }
   }
 
   inline bool Validate(const byte* pc, WasmOpcode opcode,
-                       SimdConcatOperand<true>& operand) {
-    DCHECK_EQ(wasm::kExprS8x16Concat, opcode);
-    if (operand.bytes <= 0 || operand.bytes >= kSimd128Size) {
-      error(pc_ + 2, "invalid byte amount");
+                       SimdShuffleOperand<true>& operand) {
+    unsigned lanes = GetShuffleMaskSize(opcode);
+    uint8_t max_lane = 0;
+    for (unsigned i = 0; i < lanes; i++)
+      max_lane = std::max(max_lane, operand.shuffle[i]);
+    if (operand.lanes != lanes || max_lane > 2 * lanes) {
+      error(pc_ + 2, "invalid shuffle mask");
       return false;
     } else {
       return true;
     }
   }
 
   static unsigned OpcodeLength(Decoder* decoder, const byte* pc) {
-    switch (static_cast<byte>(*pc)) {
+    WasmOpcode opcode = static_cast<WasmOpcode>(*pc);
+    switch (opcode) {
 #define DECLARE_OPCODE_CASE(name, opcode, sig) case kExpr##name:
       FOREACH_LOAD_MEM_OPCODE(DECLARE_OPCODE_CASE)
       FOREACH_STORE_MEM_OPCODE(DECLARE_OPCODE_CASE)
 #undef DECLARE_OPCODE_CASE
       {
         MemoryAccessOperand<true> operand(decoder, pc, UINT32_MAX);
         return 1 + operand.length;
       }
       case kExprBr:
       case kExprBrIf: {
@@ skipping 62 matching lines @@
 #undef DECLARE_OPCODE_CASE
           {
             return 2;
           }
 #define DECLARE_OPCODE_CASE(name, opcode, sig) case kExpr##name:
           FOREACH_SIMD_1_OPERAND_OPCODE(DECLARE_OPCODE_CASE)
 #undef DECLARE_OPCODE_CASE
           {
             return 3;
           }
+          // Shuffles contain a byte array to determine the shuffle.
+          case kExprS32x4Shuffle:
+          case kExprS16x8Shuffle:
+          case kExprS8x16Shuffle:
+            return 2 + GetShuffleMaskSize(opcode);
           default:
             decoder->error(pc, "invalid SIMD opcode");
             return 2;
         }
       }
       default:
         return 1;
     }
   }
 
@@ skipping 1022 matching lines @@
     SimdShiftOperand<true> operand(this, pc_);
     if (Validate(pc_, opcode, operand)) {
       compiler::NodeVector inputs(1, zone_);
       inputs[0] = Pop(0, ValueType::kSimd128).node;
       TFNode* node = BUILD(SimdShiftOp, opcode, operand.shift, inputs);
       Push(ValueType::kSimd128, node);
     }
     return operand.length;
   }
 
-  unsigned SimdConcatOp(WasmOpcode opcode) {
-    DCHECK_EQ(wasm::kExprS8x16Concat, opcode);
-    SimdConcatOperand<true> operand(this, pc_);
+  unsigned SimdShuffleOp(WasmOpcode opcode) {
+    SimdShuffleOperand<true> operand(this, pc_, GetShuffleMaskSize(opcode));
     if (Validate(pc_, opcode, operand)) {
       compiler::NodeVector inputs(2, zone_);
       inputs[1] = Pop(1, ValueType::kSimd128).node;
       inputs[0] = Pop(0, ValueType::kSimd128).node;
-      TFNode* node = BUILD(SimdConcatOp, operand.bytes, inputs);
+      TFNode* node =
+          BUILD(SimdShuffleOp, operand.shuffle, operand.lanes, inputs);
       Push(ValueType::kSimd128, node);
     }
-    return operand.length;
+    return operand.lanes;
   }
 
   unsigned DecodeSimdOpcode(WasmOpcode opcode) {
     unsigned len = 0;
     switch (opcode) {
       case kExprF32x4ExtractLane: {
         len = SimdExtractLane(opcode, ValueType::kFloat32);
         break;
       }
       case kExprI32x4ExtractLane:
@@ skipping 17 matching lines @@
       case kExprI32x4ShrU:
       case kExprI16x8Shl:
       case kExprI16x8ShrS:
       case kExprI16x8ShrU:
       case kExprI8x16Shl:
       case kExprI8x16ShrS:
       case kExprI8x16ShrU: {
         len = SimdShiftOp(opcode);
         break;
       }
-      case kExprS8x16Concat: {
-        len = SimdConcatOp(opcode);
+      case kExprS32x4Shuffle:
+      case kExprS16x8Shuffle:
+      case kExprS8x16Shuffle: {
+        len = SimdShuffleOp(opcode);
         break;
       }
       default: {
         FunctionSig* sig = WasmOpcodes::Signature(opcode);
         if (sig != nullptr) {
           compiler::NodeVector inputs(sig->parameter_count(), zone_);
           for (size_t i = sig->parameter_count(); i > 0; i--) {
             Value val = Pop(static_cast<int>(i - 1), sig->GetParam(i - 1));
             inputs[i - 1] = val.node;
           }
@@ skipping 661 matching lines @@
 BitVector* AnalyzeLoopAssignmentForTesting(Zone* zone, size_t num_locals,
                                            const byte* start, const byte* end) {
   Decoder decoder(start, end);
   return WasmDecoder::AnalyzeLoopAssignment(&decoder, start,
                                             static_cast<int>(num_locals), zone);
 }
 
 }  // namespace wasm
 }  // namespace internal
 }  // namespace v8